Paper
Generative Adversarial Networks
Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, Yoshua Bengio
arXiv:1406.2661 · 0▲ · stat.ML, cs.LG
View on arXiv →Premium readers get an interactive explainer for this paper — a figure you can poke at, not just read.
What it is
This is the original Generative Adversarial Nets (GAN) paper. It trains two neural networks simultaneously in a minimax game: a generator G that maps random noise to fake samples, and a discriminator D that tries to tell real training data from generated data, with both trained by ordinary backpropagation.
Why it matters
It gave practitioners a way to train generative models that produce sharp samples using only forward and backward passes, with no Markov chains, no partition function estimation, and no approximate inference at training or sampling time. Sampling is a single forward pass through the generator, which is cheap and fast compared to Boltzmann machines or GSNs that require Markov chain mixing.
Practical takeaway
This is the foundation for the entire GAN line of work (image synthesis, style transfer, super resolution). Watch for the training instability the paper itself flags: G and D must stay synchronized, and mode collapse (the authors call it the Helvetica scenario) happens when G is over-trained relative to D. The suggested fix of training G to maximize log D(G(z)) instead of minimizing log(1 - D(G(z))) is still standard practice for avoiding vanishing gradients early in training.
Key result
On Parzen window log-likelihood estimates the adversarial nets scored 225 +/- 2 on MNIST (best of the compared models: DBN, Stacked CAE, Deep GSN) and 2057 +/- 26 on TFD (second to Stacked CAE's 2110). The authors caveat that this metric has high variance and does not perform well in high dimensions, and sample quality is only judged qualitatively as competitive rather than superior.
Subscribe
Get the next issue.
Free. One email a week. Unsubscribe any time: no account, no dark patterns.